🕐 Cron Scheduling
cron.py turns Jarvis from reactive to proactive. Scheduled tasks in standard cron format with automatic error recovery.
📌 Scheduled Task Examples
- •0 9 * * 1-5 → daily pending items report at 9 a.m. (weekdays)
- •*/30 * * * * → check for important emails every 30min
- •0 18 * * 5 → weekly summary every Friday at 6 p.m.
- •0 0 * * 0 → back up memory.db every Sunday at midnight
- •*/5 * * * * → heartbeat for critical services every 5min
💡 Practical Tip
Avoid schedules that run too frequently with expensive models. A heartbeat every 5min with GPT-4o costs ~R$50/day. Use inexpensive models for periodic checks.
💓 Heartbeat — Active Monitoring
Heartbeat is Jarvis’s nervous system. Configurable periodic checks that proactively alert you when something is out of the ordinary.
📌 What the Heartbeat Monitors
- •Critical service APIs: status code != 200 → alert
- •Memory usage: memory.db > 100MB → compact
- •Daily API cost: > threshold → alert and throttle
- •Important files: did the hash change? → possible compromise
- •Jarvis's own uptime: response timeout → auto-restart
💡 Practical Tip
Set up heartbeats at multiple intervals: fast (5min) for critical services, slow (1h) for non-urgent checks. Balance cost vs. responsiveness.
🔗 Webhooks with n8n and Zapier
Webhooks connect Jarvis to the existing automation ecosystem. n8n, Zapier, Make.com can trigger Jarvis actions via HTTP.
📌 Webhook Integration Flow
- •n8n detects a new lead in the CRM → POST to /webhook/new-lead
- •Jarvis receives an event, analyzes it with an LLM, and generates a personalized email
- •Jarvis calls GmailTool to send an email to the lead
- •Jarvis updates the CRM via API with the action result
- •Audit log records the entire chain
💡 Practical Tip
Always validate the webhook’s HMAC signature before processing. Unauthenticated webhooks are vulnerable to replay attacks and spam.
🤖 Background Sub-Agents
Sub-agents enable true parallelism. The main Jarvis creates sub-agents for long-running tasks without blocking the main conversation.
📌 Sub-Agent Use Cases
- •Large repository analysis: a sub-agent scans and indexes while the main agent talks
- •Extensive web research: 3 sub-agents research in parallel, the main agent aggregates
- •Report compilation: sub-agent collects data, primary agent formats and sends it
- •Continuous monitoring: sub-agent monitors in a loop, main agent receives alerts
- •Deploy pipeline: a sub-agent executes steps, the main agent reports progress
💡 Practical Tip
Limit the number of simultaneous sub-agents (recommended: max 5). Each sub-agent consumes tokens and incurs costs. Use an agent pool for reuse.
🤝 TinyClaw Multi-Agents
TinyClaw is the standard for INTELECTO agent teams. Specialized agents collaborate passing work among themselves with defined roles.
📌 Common Roles in TinyClaw
- •@coder: generates code and tests based on the spec
- •@reviewer: analyzes generated code for bugs and style
- •@writer: documents the code and generates a changelog
- •@deployer: runs deploy scripts with approval
- •@orchestrator: coordinates work among the other agents
💡 Practical Tip
Each TinyClaw agent has a specialized SOUL.md. @reviewer is deliberately critical; @coder focuses on functionality; @writer prioritizes clarity.
📊 System Observability
Without observability, optimization is guesswork. metrics.py collects latency, cost, errors, and real-time tool usage.
📌 Collected Metrics
- •Latency per request: p50, p95, p99 in ms
- •Tokens per request: input + output + cost in USD
- •Tool calls: frequency and success rate per tool
- •Error rate: by error type and provider
- •Agent loop depth: distribution of rounds used per conversation
💡 Practical Tip
Export metrics to a dashboard via the /metrics endpoint. Grafana + Prometheus, or just a simple endpoint that Jarvis reads when asked.
✅ Module 4.3 Summary
Next:
Track 5: Architectures